This project provides a simulated world populated by AI agents powered by large language models (LLMs), building upon Stanford's "Generative Agents" research. It's designed for researchers and developers interested in creating and observing complex emergent behaviors in AI-driven simulations, offering a Chinese-language interface and mobile compatibility.

How It Works

The simulation environment is based on the Generative Agents framework, which uses LLMs to imbue agents with memory, planning capabilities, and social interactions. This version enhances the original by offering a Chinese interface, optimized performance, support for local LLMs, and adaptation for mobile platforms like HarmonyOS. It also includes bug fixes for stability issues found in the original demo.

Quick Start & Requirements

• Installation: pip install -r requirements.txt (or with Tsinghua mirror: pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple)
• Prerequisites: Python 3.9.12 (3.8+ recommended). Requires configuration of API keys or local LLM paths in reverie/backend_server/utils.py.
• Setup: Basic setup involves installing dependencies and configuring LLM access. Running the simulation requires starting two services: a Django frontend server and a backend agent service.
• Resources: An RTX 4070 GPU can run a 7B LLAMA2 model at medium speed; API calls are faster but incur costs.
• Links: Generative Agents Paper

Highlighted Details

• Supports Chinese interface and local LLM integration (e.g., LLaMa2, GPT4ALL, Falcon).
• Optimized for mobile (HarmonyOS) and includes bug fixes for common crashes.
• Allows for customization of agent backgrounds and simulation parameters.
• Includes a replay mechanism for saved simulation states.

Maintenance & Community

The project is based on Stanford's Generative Agents research, with authors listed as Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Further community or maintenance details are not specified in the README.

Licensing & Compatibility

The README does not explicitly state a license. Given its foundation on Stanford's research, users should verify licensing terms for commercial or derivative use.

Limitations & Caveats

The simulation runs with a delay, generating steps offline before displaying them. The default model uses OpenAI, which can be costly with many agents or simulation steps. Network instability may cause API call failures, though a real-time save feature is included to mitigate data loss. Replay functionality for full simulations requires manual execution of a Python script (compress_sim_storage.py).